Understanding how often a model fabricates traits for nonexistent organisms is crucial before relying on its summaries for real-world microbial annotation
To probe where certainty ends and storytelling begins, we built a library of 200 wholly invented strain names, arranged along a realism gradient. At the playful end sit English mash-ups such as Crimson Horizon or Silver Pine—labels no taxonomist would buy. At the serious end are Latin-looking binomials like Luminaricella splendens that follow every rule of bacterial nomenclature. By asking the same set of questions across this spectrum, we can see exactly when an LLM's confidence tips into hallucination.